- Signal fusionTactile, force, and proprioceptive signals often matter more than bigger vision backbones.
- Recovery behaviorModels must handle retries, slip, and micro-adjustments.
- Latency toleranceReal control loops need predictable inference behavior.
Models for contact-rich manipulation
Contact-rich tasks often fail because the model never sees the right signals, not because the network is too small.
For contact-rich work, better supervision and signal design usually beat chasing the largest general model.